All Watched Over By Machines Of Loving Grace
I like to think (and
the sooner the better!)
of a cybernetic meadow
where mammals and computers
live together in mutually
programming harmony
like pure water
touching clear sky.
I like to think
(right now, please!)
of a cybernetic forest
filled with pines and electronics
where deer stroll peacefully
past computers
as if they were flowers
with spinning blossoms.
I like to think
(it has to be!)
of a cybernetic ecology
where we are free of our labors
and joined back to nature,
returned to our mammal
brothers and sisters,
and all watched over
by machines of loving grace.
Richard Brautigan
tl;dr
Local communities have shown that they can manage shared resources sustainably by governing themselves. Self-governance breaks down at larger scales, though, because people have limited attention and coordination gets harder as groups grow. In this essay I propose studying "Regenerative Intelligence" (RI), the theory and practice of designing sociotechnical intelligent systems that preserve or enhance social capital, trust and agency while scaling governance capabilities. Examples of RI in practice include (1) paying communities to collect environmental data (Conservation Data Income), (2) impact certificates (hypercerts), (3) AI assistants that help Indigenous communities share knowledge and help youth from the Global Majority take part in climate negotiations, and (4) our work with Indigenous scientists, which helped us win the XPRIZE Rainforest.
Introduction
Can we scale human cooperation through trustworthy machines?
I have been thinking about this question for years1. At GainForest.Earth, the non-profit I co-founded, we study how people and machines can cooperate on what is arguably our planet's worst coordination failure, the climate and biodiversity emergency.
Garrett Hardin described the economics behind this crisis in his famous essay Tragedy of the Commons2. People exploit limited shared resources without limit, and that drives our environmental emergencies. Global deforestation is the clearest example of this race to collective ruin. Demand for beef, palm oil and soy keeps putting short-term profit ahead of the forest. The results are severe. Half the world's forests are gone, their loss has released a tenth of global anthropogenic emissions, and hundreds of species go extinct every day. The economics show how badly global coordination has failed. Halting deforestation would cost about 200 billion USD a year, a fraction of the estimated 35 trillion USD in ecosystem services at risk. Yet international funding commitments are stuck below 10% of what's needed. Some scientists suggest we may already have crossed planetary tipping points, so we need global action now.
Fortunately, people are less grim than Hardin makes them out to be. Nobel laureate Elinor Ostrom showed that local communities can manage shared resources sustainably by governing themselves. In her book Governing the Commons she identified eight design principles3 behind fair and lasting management of the commons, and she backed them with successful examples that span centuries.
Clear boundaries
Local implementation: Clear physical borders with known community members
Real-world examples: Maine lobster fishing zones marked by harbors; Japanese iriai forests bounded by village territories
Local rules
Local implementation: Traditional practices adapted to local conditions
Real-world examples: Swiss Alpine pasture rotation by seasons; Spanish huertas water schedules adapted to local soil types
Collective choice
Local implementation: Direct participation through community meetings
Real-world examples: New England town halls managing common lands; Mexican ejido assemblies governing communal forests
Monitoring
Local implementation: Community patrols and mutual observation
Real-world examples: Nepali community forest guards; Turkish fishers watching designated coastal zones
Graduated sanctions
Local implementation: Flexible penalties set by the community
Real-world examples: Tohono O'odham graduated fines for water violations; Japanese village penalties for forest overuse
Conflict resolution
Local implementation: Local mediators and traditional councils
Real-world examples: Valencia's Water Court resolving irrigation disputes; Korean village elders mediating water conflicts
External recognition
Local implementation: Official acceptance of traditional rights
Real-world examples: Mexican recognition of ejido lands; Nepal's legal status for community forests
Nested systems
Local implementation: Traditional hierarchies linking local to regional
Real-world examples: Swiss cheese cooperatives connecting to markets; Balinese subak temples organizing from village to region
Table 1: Examples of successful self-governance of the commons from Ostrom's work. They show her principles working in very different cultures.
Ostrom's work showed that communities can organize themselves and govern resources sustainably when they have the autonomy, tools and institutional support to do so. But the tragedy Hardin described comes back at larger scales. There her principles get much harder to apply, as we see with climate change, ocean pollution and the deforestation of the Amazon.
Still, people are inventive. Web3 lets us design economic incentives with a precision and speed that wasn't possible before, down to a single bit. Add the likely arrival of powerful AI4, and computed incentives together with machine intelligence could change how well humans cooperate. If we guide this combination carefully, it could let us manage natural and digital commons at scale. We call this emerging discipline regenerative intelligence (RI).
In this essay I explore RI and how it could scale Ostrom's eight principles in the intelligent age. I draw on our team's early pilots with 30 communities around the world, from the Philippines to the Amazon rainforest, and on our recent XPRIZE Rainforest win. I look at how RI could change the way we govern natural commons. I also set out a theory of change for how RI could address the biodiversity and climate emergencies by turning humanity's race to the bottom into a collective race to the top.
Basic assumptions and definitions
Ostrom's principles across different scales
Ostrom drew her eight principles from decades of research around the world. They have shaped local policy and management systems, but they are hard to apply at larger scales. Table 2 compares the two.
Clear boundaries
Why it works locally: Geography draws clear borders, and everyone knows who's in the community
Why it's hard at scale: Resources like air and oceans cross borders, and users are anonymous and numerous
Local rules
Why it works locally: Communities can adapt rules from direct experience and local knowledge
Why it's hard at scale: Different regions need different rules, and one-size-fits-all policies often fail
Collective choice
Why it works locally: Face-to-face meetings allow direct participation and quick consensus
Why it's hard at scale: There are too many stakeholders to coordinate, and representation gets complicated
Monitoring
Why it works locally: Daily contact makes violations easy to spot, and monitors are known community members
Why it's hard at scale: The areas are vast, violations are hard to detect and monitoring is expensive
Graduated sanctions
Why it works locally: Everyone knows each other, so penalties can fit the circumstances
Why it's hard at scale: Violators are hard to identify, and enforcement across jurisdictions is complex
Conflict resolution
Why it works locally: Local mediators and councils make it quick, informal and cheap
Why it's hard at scale: Several jurisdictions are involved, legal costs are high and cultures differ
External recognition
Why it works locally: Traditional rights and authority are clear and respected locally
Why it's hard at scale: Legal systems may conflict, and international recognition is complicated
Nested systems
Why it works locally: Hierarchies grow naturally out of local relationships
Why it's hard at scale: Coordinating many levels is complex, and interests compete
Table 2: Ostrom's principles at local and larger scales, and why the traditional mechanisms are hard to scale up.
What makes Ostrom's principles work locally is direct relationships, shared context and face-to-face meetings. Those same things stop them from working at larger scales.
There are several ways to explain why. One comes from social anthropology. Dunbar's number is the cognitive limit on how many stable relationships one person can keep, and it is thought to be around 1505. When a community stays within that limit, governance works because everyone can keep real relationships and trust each other directly. Past that limit, complexity grows non-linearly. A group of n people has n(n-1)/2 possible relationships. A community of 150 has 11,175 of them to keep track of. A community of 1,500 has 1,124,250.
Attention is all we need (and expensive)
Another explanation comes from attention economics, first described by Herbert A. Simon, who won both the Nobel Prize and the Turing Award6. Human attention is finite, and decisions need it. Imagine that each person has a fixed number of "attention tokens" to spend on governance. Locally, the tokens go a long way. People keep an eye on familiar places during their daily routine, settle conflicts in quick informal chats and make collective decisions in regular face-to-face meetings. As systems grow, the demands on attention multiply fast. A fisher can watch the local coastline for violations while working, but can't also track industrial fleets across the oceans. A village elder can mediate a dispute between neighbors they know, but can't keep the context of thousands of cases across several jurisdictions. Community members can take a real part in a local assembly, but can't follow everything that goes into global climate policy. This is why simply scaling up local governance fails. We run out of attention tokens long before we can apply the same human judgment and oversight at a larger scale.
Attention gets even scarcer once you look at how people think. Nobel laureate Daniel Kahneman7 describes two systems we use to make judgments, a fast, intuitive "System 1" and a slower, rational "System 2". System 1 makes mistakes because of cognitive biases like anchoring (leaning too hard on first impressions), availability bias (thinking events are more likely when they're easy to remember) and confirmation bias (looking for information that confirms what we already believe). It also takes mental shortcuts called heuristics, such as the representativeness heuristic (judging by stereotype instead of base rates) and the affect heuristic (letting feelings drive the decision). System 2 can catch these mistakes through careful analysis, but it takes a lot of mental effort.
Other scaling problems
Attention is only one part of the puzzle. As governance grows, other problems show up. Local and higher-level decision makers know different things. Power dynamics and competing interests appear. The coordination problems themselves also change. Managing a local commons mostly means balancing the needs of community members everyone knows. Climate change means negotiating between actors with very different interests, capabilities and access to information. Larger problems are different problems, which is why expanding local governance mechanisms usually fails.
Three ways to scale governance
To scale governance we need to expand several capacities that limit human coordination today. They fall into three groups that complement each other:
1. Intelligence augmentation: amplifying human capabilities and judgment
2. Intelligent automation: handing routine, data-driven work to machine protocols
3. Capacity building: adding skills, capabilities and resources
Ostrom's principles are different enough from each other that we need all three, as Table 3 shows.
Clear boundaries
What needs improvement?: Mostly technology
Why?: Digital tools can verify who belongs to the community and control access. This is mainly a technical problem, and computers handle it well.
Local rules
What needs improvement?: Mostly community skills
Why?: Rules have to reflect local culture and wisdom. That takes human understanding, which can't be fully automated.
Collective choice
What needs improvement?: Both technology and community skills
Why?: Voting can be automated, but complex decisions need good digital tools and human judgment.
Monitoring
What needs improvement?: Mostly technology
Why?: Sensors and pattern-detection software can watch over resources with little human intervention.
Graduated sanctions
What needs improvement?: Both technology and community skills
Why?: Fair enforcement needs automated tracking and a human sense of context.
Conflict resolution
What needs improvement?: Mostly community skills
Why?: Resolving conflicts takes a deep understanding of relationships and local context, which machines can't replicate well.
External recognition
What needs improvement?: Both technology and community skills
Why?: Proving authority takes technical systems, such as digital certificates, and social acceptance.
Nested systems
What needs improvement?: Mostly technology
Why?: Automated systems and smart contracts can handle much of the coordination between levels of organization.
Table 3: How each of Ostrom's principles can be improved. Different governance problems need different mixes of technology and social skills.
Some principles are good candidates for intelligent automation, like monitoring at scale. Others need intelligence augmentation, like collective choice, or depend mostly on building human capacity, like conflict resolution. The reason is that many governance tasks need local knowledge, wisdom and social context, especially when people decide with incomplete information (adapting local rules) or work across institutions (coordinating policy). Every principle can still benefit from intelligent infrastructure that helps manage data and make decisions, as long as it keeps the overhead of working with machines low.
Scaling governance well means building new sociotechnical capabilities that give people more agency while they tackle these problems. Automation is one part of that. Technology can get us past some limits, but incomplete information and irreducible uncertainty will probably always constrain us. That brings us to the core hypothesis of regenerative intelligence.
Defining regenerative intelligence
With the scaling problems and the possible technical fixes laid out, we can define regenerative intelligence more precisely. It sits where traditional governance wisdom meets modern technology.
Regenerative Intelligence (RI)
The theory and practice of designing sociotechnical intelligent systems that preserve or enhance social capital, trust and agency while scaling governance capabilities
Each part of the definition answers one of the scaling problems above:
"Sociotechnical" because technology alone won't work, and neither will social fixes alone
"Intelligent systems" because the focus is data-driven, intelligent infrastructure
"Preserve or enhance social capital" because we don't want to trade community bonds for efficiency
"Trust and agency" because local autonomy and self-governance stay central, as in Ostrom's work
"Scaling governance capabilities" because the goal is effective governance beyond the local level
That gives RI its fundamental problem:
The fundamental problem of RI:
How can we design sociotechnical intelligent systems that:
• Augment human governance capabilities across scales
• Preserve or enhance human agency
• Maintain legitimate governance and trust
• Align machines with the values of diverse communities
RI differs from fields like human-computer interaction and mechanism design because it looks at three things together: scaling human cooperation, keeping trust and aligning technology with social values. Traditional approaches often trade social dynamics for efficiency or stay with small-scale interactions. RI takes on the harder problem of keeping governance effective as systems grow past natural human limits. It combines technology with social wisdom to work on global problems like climate change. The aim is to add to our shared capacity to coordinate and govern common resources, with digital trust built in and human judgment kept in place.
GainForest's framework to scale Ostrom's principles
Building RI is hard. The systems have to scale governance without taking agency away from people or losing their trust. At GainForest we're building the foundations through what we call the self-improving sociotechnical loop, or SISL (pronounced "zizzle"), a framework for improving governance capabilities cycle after cycle.
SISL has five stages that repeat:
- 1.
Gather, collaborate & govern: communities govern and work together directly with digital tools
- 2.
Improve data: these interactions and local data collection improve our infrastructure and understanding
- 3.
Improve tools: better data lets us build better tools and systems
- 4.
Capacity building: better tools support capacity building
- 5.
Utility feedback: every cycle gives all stakeholders quick feedback on the measured utility
At every stage we reflect, align and improve, and that feeds back into the other stages. The goal is more utility with each cycle, so governance gets more effective and scalable while people keep their agency and social capital.
Here is how GainForest maps SISL onto technology we can use today:
Decentralized infrastructure, such as splits contracts, guilds and DAOs, lets us build community across local commons regardless of borders and gives communities access to governance tools.
Better data infrastructure, such as hypercerts and data markets, helps groups make sense of information together and allocate resources.
Better data makes better tools for governance and coordination possible.
Co-designing powerful AI systems as "machine classmates" instead of replacements builds human capacity while protecting agency, social capital and, above all, trust.
Collaborate and govern with decentralized infrastructure
As I write this, our non-profit supports 30 grassroots communities around the world. We grew organically, mostly through peer recommendations, and you can see that in where our communities are. Thanks to our local community champions, regional hubs have formed in South-East Asia, East Africa and South America.
Not every community uses our full technology stack. Most receive funding through a mechanism we call Conservation Data Income, which I describe below under "Lessons from real-world deployment". Many join for our monthly skill-sharing lessons. Some use the whole stack and have improved their governance a lot through extra funds, capacity building and monitoring. What they all share is that they use cryptocurrency.
Cryptocurrency has been invaluable for working across borders8. The crypto community, especially around Ethereum, has built some of the best tools in the world for decentralized governance. One reason is that no single entity owns a blockchain. Node operators around the world run it together.
Global Majority9 countries need reliable financial systems, and that gives communities a chance to adopt new technology faster than the Global North. M-PESA in Kenya is a good example. It's a centralized digital money system that most Kenyans use every day in place of more expensive and slower credit cards. Since it launched in 2007, M-PESA has grown to more than 15 million transactions a day and is the infrastructure behind 59% of Kenya's GDP. It shows how leapfrogging to new technology can change an entire economy. Statista estimates that 653 million people owned cryptocurrency in 2024, about 8% of the world, and many of them live in the Global Majority.
So GainForest has always coordinated through a shared smart contract that releases funds once certain criteria are met, such as avoided deforestation10. A public wallet address is your immutable digital identity. It gets registered, receives funds and builds up a reputation. Public wallets are a remarkable tool for Ostrom's first principle, clear boundaries. They make fine-grained access control easy, so an address can gain, transfer or lose access to digital resources and privileges in a fraction of a second.
We use a decentralized form of guild, which we call the Nature Guild, to help our communities move to the decentralized web. Guilds go back to the Sumerian city of Uruk around 3000 BCE. They started as temple-based craft workshops and later became associations of artisans and merchants who oversaw their craft or trade in a territory. They had a shared treasury and member benefits, and they governed themselves, with members deciding who could belong. Decentralized guilds copy these systems with tighter feedback loops. Members, identified by their wallet addresses, can propose and vote on adding or removing other members. Donations to the guild go to a smart contract, which works like an automated bank account. It splits the money according to weights that each member sets every quarter, based on criteria such as activity in the guild, impact on nature projects or need.
The strength of a decentralized guild is that you can extend it piece by piece. Members get a members-only digital forum to talk in, and quarterly fund releases with programmatic weights keep everyone accountable. Because every member and the guild itself have a unique digital identity, we can plug in better data, tools and regenerative intelligence to improve the guild's decisions.
Taking part in decentralized governance also gives grassroots communities a voice in how technology and resources are allocated in the crypto space itself. That matters, because the treasuries of the five largest DAOs (Decentralized Autonomous Organizations) together hold about $25 billion. Our members used their skills as a way into these communities, and in the past that gave them access to funding that had been hard to get through traditional channels.
Better data with hypercerts and data markets
Frontline communities around the world lack the data they need to make good decisions. Of all the biodiversity data ever collected, 82.7% comes from North America and Europe11. Only 0.5% comes from the Amazon rainforest, where scientists estimate most of the world's biodiversity lives. The Amazon is home to an estimated 15% of all known species and possibly millions more that no one has described yet. At GainForest we see this gap as an opportunity. Local communities can create a lot of value by collecting data together.
The barriers that created the gap also make it hard to close. Communities lack infrastructure, digital skills training and access to computers and hardware. We follow a cosmo-local approach, where knowledge flows globally and production stays local. Once a community is connected to the network, digital knowledge can spread within it. That's why internet access, such as SpaceX's Starlink, and basic computers are the first investments and milestones when we bring a community on board. Starlink grew from 100,000 users in 2021 to more than 4.6 million customers in over 118 countries in 2024. The obstacles are big, but communities are adopting these technologies faster every year, and that excites us. Mobile internet keeps spreading too. Today 57% of the world's population, 4.6 billion people, use mobile internet on their own device, so digital governance of natural commons is becoming realistic.
Connectivity makes it possible to fix environmental data collection, but infrastructure alone isn't enough. We need a system that handles several problems at once. It has to standardize data collection across very different communities, keep data quality high, give people a reason to keep participating and pay communities fairly for their work. It has to scale globally and still adapt locally. It also needs clear verification so that data collectors and data users can trust each other. Traditional approaches, which usually rely on centralized databases or informal arrangements, have struggled to do all of this at once. We need new technology combined with the right economic incentives, and this is where ecological hypercerts come in.
Ecological hypercerts are blockchain-based impact certificates. They follow a structured data standard with three parts: identity (who made the impact), formalized claims (what impact was achieved) and evidence (proof of the impact). Each hypercert is an immutable record that links to its evidence and to public datasets. On top of this, a regenerative intelligence can track reputation, run rigorous evaluations and monitor impact over time, and communities can improve and extend their work based on the feedback. The appeal of hypercerts is that they're simple. They give impact data a clear, standard interface only when it's needed. Compared with traditional, exhaustive methodologies, that makes it much cheaper for communities and machines to collect, share and connect information. This shared layer is what lets data-driven decisions and joint evaluation scale up positive impact.
Most importantly, hypercerts are digital proof of impact, and that creates a market for ecological data. Carbon credits created a data market for measuring and delivering carbon reductions12. Hypercerts do the same for collecting and verifying ecological data. Communities have a reason to collect and validate data because hypercerts attract donors who want to fund verified impact. This can become a self-reinforcing cycle. Better data means more accurate impact measurement, which attracts more donors, who pay for more conservation work and data collection. We believe the benefits go beyond funding, since communities also get good evaluation tools and personalized reports on their impact.
Decentralizing marketplaces for environmental data
To build fair data markets that make good use of hypercerts, we first need to understand the economics of data itself. Its properties shape how we can design markets that serve both local communities and global environmental goals.
Data is non-rival13. Many people can use it at the same time without making it worse. The data economy has a paradox, though. Very large datasets need so much infrastructure, storage and processing that these needs become barriers to entry. The barriers create artificial scarcity, and data often ends up concentrated in a few hands.
Datasets also tend to become more valuable as they grow, because of network effects. Each new contributor and data point makes the insights more useful. This matters a lot for ecological data, where combining different datasets can reveal patterns in ecosystem health and biodiversity that no single dataset shows. The cost of producing more data falls with scale, but the upfront infrastructure costs are high, and they can shut out smaller participants and the local perspectives they bring.
Data is also an experience good. You often don't know its real value until you've used and analyzed it. That creates information gaps between the people who collect data, like local communities, and the people who buy or use it. Environmental data markets need quality assurance and standards before anyone can trust them.
At GainForest we're working on these problems with LUCA, a decentralized marketplace that connects hypercerts with environmental data providers. Ecological hypercerts create demand for evidence, and LUCA lets hypercert creators, buyers and evaluators query and attest the data available for any geographic area. A data router connects users to a decentralized coalition of providers, including local communities, NGOs, scientific databases and private companies. Users get broad access to environmental data, and providers keep local ownership and control. To manage access and keep the economics sustainable, LUCA charges for queries in tokens. This creates controlled scarcity and pays data providers fairly. The design tries to balance access against quality, so communities can benefit from network effects and still control their data.
We believe that combining the hypercert standard with decentralized markets can open environmental data to everyone, decentralize impact attestations around each hypercert and build trust in formal impact claims. The value then flows back to local data collectors, verifiers and grassroots communities.
Better tools with agentic function calls
Our communities work on different projects, but they often need the same things, such as monitoring at scale and biodiversity assessments. At GainForest we translate these needs into automated workflows. Counting trees, for example, takes a series of steps. Community members fly drones and upload the images. Software stitches the images together, machine learning classifiers identify the species, our web interface publishes the results, and the findings go back to the community. Tools like these standardize how information gets processed and let it scale.
We help build these tools, but someone also has to run them at scale. One answer is automated workflows run by AI. AI agents learn a community's requirements from examples, pick tools from our shared inventory, plan the workflow and carry it out. This works well because community members get quick feedback at any time, and all they need is a chat interface. We imagine an ecosystem built around a shared tool inventory, run by a regenerative intelligence that plans and supports community workflows. It talks with communities through prompts and replies and picks the right tools for each task.
To get there we need more data and close collaboration between people who hold local knowledge and people who design AI. We need to know which workflows communities use most, what typical conversations look like and how to make tools easier for machines to use.
The tools we've identified so far include translation (important for climate negotiations), drone image analysis, bioacoustic monitoring and writing hypercerts. The central idea is a registry of community-owned digital commons, tools that the communities maintain and develop together. Good tools also add to existing data and generate synthetic data, which we can use to train and improve the regenerative intelligence. Each round of improvement feeds the next one, and communities gain more control along the way.
Capacity building with human-machine co-learning
A common mistake in the tech industry is building a solution before finding a real problem. It happens when the people the technology is meant for are left out of building it. Technology for the Global Majority suffers from this a lot. Frontline communities are often hard to reach, and from a Silicon Valley point of view there's no business case to pitch to venture capitalists.
What we've seen is simple: Talent is equally distributed but opportunity is not.
So we don't outsource technology development to wealthy nations. We build capacity inside communities. That creates a skilled local digital workforce, and each round of progress makes the next one easier. When the problem-solvers live with the problems, the technology reflects more worldviews and experiences. Technology always carries the views of the people who build it, so including more views makes solutions more inclusive and more resilient.
A skilled local workforce pays off now and later. Community members we trained have gone on to start their own ventures, become financially independent and put money back into their communities and the commons. Given the opportunity, local talent drives real technical and social change. The Global Majority makes up 80% of the world's population. It is also by far the youngest part of it, and eager to build the world of tomorrow.
How do we scale capacity building? We think the answer is fair collaboration between people and machines. We already learn with computers every day, from calculators to search engines. Modern AI adds interactive learning that adapts to the learner. At GainForest we've changed how we think about AI in education. Instead of teachers or tools, we see AI systems as "machine classmates", learning partners who discover and master concepts alongside people. Human classmates offer different perspectives, ask clarifying questions and help work through problems. AI systems can do the same while adapting to each community's context and needs.
The classmate idea works because it keeps people in charge of their own learning. Classmates learn together as equals, without a teacher-student hierarchy. We've found that when communities treat AI systems as fellow learners instead of authorities, they question the technology and its suggestions more, and they fit it into their own knowledge systems with more care.
Community-aligned regenerative intelligence
As our relationship with regenerative intelligence grows, its values have to match those of each community. That builds trust and prevents misalignment. But how do you program such complex machines together with communities that are new to digital technology?
It turns out to be quite doable. Building machine learning and AI models has never been easier. As machines get more capable, we can work with them at higher levels of abstraction. We went from punch cards in the 1950s and low-level assembly code to higher-level languages like Python, Rust and Go. Today we can program computers in plain human language.
Constitutional AI14 is a good example. It aligns AI systems through explicit principles, which the AI company Anthropic calls a "constitution". Direct, principle-based instructions like these often work better than learning from demonstrations, where the AI has to work out the principles from examples of right and wrong behavior. We've seen this in our co-design workshops. Communities helped write the system prompt for our first regenerative intelligence, Taina, and gave it a personality that reflects their values and earns their trust.
Co-design and value-aligned constitutions are only the start. AI models keep getting more capable while needing less compute, and powerful models now run on local machines, which was hard to imagine even a year ago. Self-hosted models give communities real data sovereignty. They decide which information to share with the wider commons and which to keep private. When GainForest helped an Indigenous community partner host their first local model, people found uses nobody had planned. They archived traditional wisdom, and some used the model to preserve and teach languages that were close to being forgotten.
Looking ahead, we imagine many regenerative intelligences speaking different dialects, each guided by its own community. Each community chooses what data to share and what to keep private, which allows many different uses while respecting its autonomy. This keeps AI development grounded in community values and needs, and it helps preserve cultural knowledge.
Lessons from real-world deployment
RI is still in its infancy, but pilot projects by GainForest and others already show real-world impact. Running the SISL framework with many different communities taught us that one of the hardest parts of scaling regenerative governance is funding that lasts and supports communities as they grow. Every stage of the loop, from the first data collection to advanced AI, needs money to keep going and to build capacity.
Funding limits how far Ostrom's principles can scale. Local communities often have the knowledge and commitment to manage resources well but lack the money for full monitoring systems or for joining wider governance networks. Our answer is a tiered approach that grows with each community. First I describe how data-centric funding gives local communities access to four levels of funding that match their data maturity. Then I introduce Conservation Data Income (CDI), GainForest's RI-augmented version of universal basic income, which solves the cold start problem by paying for important data and infrastructure. After that I look at how CDI can expand to teach local communities downstream analytics and data skills.
As communities move up these funding levels, they can use more advanced forms of regenerative intelligence. That brings us to our first RI agents. Taina came out of co-design sessions with communities in the Amazon rainforest. Her sister agent Polly supports climate negotiators from the Global Majority in international climate talks. Finally, I describe how funding, capacity building and AI assistance came together when GainForest won the XPRIZE Rainforest.
Four levels of data-centric nature funding
Small grassroots environmental communities often struggle to find lasting funding and income for conservation. Carbon credit markets look like an opportunity, but certification through registries like Verra or Gold Standard usually costs hundreds of thousands of dollars upfront. That often pushes small projects into depending on outside project developers, who act as middlemen and take a large share of the revenue.
We believe the SISL framework opens a more accessible, step-by-step path built on data collection and developing regenerative intelligence. Through SISL, communities can reach four levels of funding that match how mature their operations are, while they build capacity and financial stability.
0. No data infrastructure
Focus: Building essential data infrastructure and core skills
Funding sources: Conservation Data Income
Market size: Investment return of the Nature Guild Principal Fund
Requirements to unlock: No requirements
1. Data collection
Focus: Field measurements, species monitoring, community impact tracking
Funding sources: Impact philanthropy, retroactive funding platforms, seed grants
Market size: $5 billion in climate philanthropy a year
Requirements to unlock: An established data collection method with at least 6 months of consistent field data, a documented community participation process and basic digital infrastructure for storing data
2. Analytics
Focus: Geospatial analysis, ecological modeling, impact assessment
Funding sources: Research grants, corporate partnerships, monitoring contracts, data marketplace
Market size: $17.9 billion environmental monitoring market
Requirements to unlock: A 12-month data collection track record, partnerships with research institutions, standardized analysis protocols, a trained local team and a data validation system
3. Market integration
Focus: Carbon sequestration, biodiversity metrics, ecosystem services
Funding sources: Carbon markets, compliance mechanisms, ESG investments
Market size: $40 billion voluntary carbon market projected by 2030
Requirements to unlock: At least 24 months of validated data, third-party verification, established baseline measurements, permanent monitoring infrastructure and certified methodologies
Table 4: Four levels of funding that grassroots communities can unlock with regenerative intelligence.
Each level requires new capabilities. Organizations without data collection experience usually start at Level 0. At Level 1 they have to show consistent data collection and community engagement, which means regular monitoring schedules, trained local teams and basic quality control. Moving to Level 2 takes stronger analytics, often through partnerships with universities or technology providers. Organizations have to show they can collect data and also turn it into insights that guide their conservation strategy.
Reaching Level 3 is the biggest milestone. Organizations need solid data infrastructure and a proven ability to run long-term monitoring, including permanent sampling plots, advanced verification protocols and relationships with certification bodies.
Level 0 funding usually ranges from USD 100 to USD 1,000 per project, depending on the Nature Guild Fund's investment returns. At GainForest we use new funding mechanisms like Conservation Data Income and SINDA to get communities past the cold start problem.
Level 1 funding usually ranges from USD 10,000 to USD 50,000 per project, and environmental philanthropy grows steadily at 12% a year15. Retroactive funding has opened new doors, with web3 funding platforms like Gitcoin and Optimism that reward good data. This level is an important entry point for grassroots NGOs. It gives them the resources to set up basic data collection.
Level 2 targets the environmental monitoring market, which is growing because companies and governments want high-quality ecological data and analysis16. Organizations that succeed at this level usually raise between USD 50,000 and USD 250,000 a year through a mix of grants and service contracts.
Level 3 has the largest funding potential and needs the largest investment. Successful projects at this level can earn more than USD 500,000 a year, but initial certification usually costs between USD 250,000 and USD 1,000,000. McKinsey projects that a gigaton-scale carbon removal market could reach USD 1.2 trillion by 2050, with nature-based solutions playing a big part17.
The framework only works if local communities keep their data quality high the whole way. Investing early in good data collection and management pays off as organizations climb the funding ladder, especially when they go for impact certification and market-based mechanisms. We believe this also helps organizations earn the trust of funders and other stakeholders, because they can show clear progress in what they can do and how they measure their impact.
Conservation Data Income (CDI)
GIF 1: A member of Toca do Tatu in the Amazon rainforest receives CDI for local data collection and pays at a local supermarket with a crypto wallet
Conservation Data Income (CDI) is the entry point to this framework. We designed it for Level 0 projects, and it builds capacity for the levels above. CDI is GainForest's main funding mechanism and a more dynamic alternative to universal basic income (UBI). UBI is unconditional. CDI ties payments to the quality and quantity of the environmental data people collect, which creates deliberate feedback loops. The micropayments, currently paid from our philanthropic endowment fund, are meant to start the SISL loop. CDI rewards three things: building digital infrastructure for data collection, setting up wallet addresses for identity and governance, and continuous capacity building within communities.
The mechanism is simple. As communities get connected and start collecting high-quality environmental data, the Nature Guild pays them in stablecoins18 in proportion to their data quality and their participation in governance. That starts the self-improving cycle. Participants want to improve how they collect data, improve their tools and grow their skills, until they can apply for Level 1 funding such as grants and donations.
Image 1: GainForest team members help local NGOs deploy AudioMoth sensors in the Southern Philippines
In CDI, the community sets the price of data itself. Every year the community votes for a data council19 that governs data ownership, policy and pricing, for example 0.01 USD per minute of bioacoustic recordings and 0.05 USD per MB of drone imagery. GainForest has run CDI in some of the most remote places in the world, including deep in the Amazon rainforest and in the Southern Philippines. There, CDI has paid for important digital infrastructure, including monthly Starlink subscriptions for communities and local NGOs, so they can take part in the digital commons. So far CDI has paid out more than 30,000 USD to communities in the Global Majority.
Towards sustainable income with nature data and AI (SINDA)
CDI covers basic data collection, but we learned that communities need more to climb the value ladder of the data economy. Data collection is the foundation of modern AI, yet it sits at the bottom of the data economy pyramid, and the work is known for exploitative conditions, especially for workers in the Global Majority. Big AI labeling companies like Scale.ai and Sama.ai rely heavily on cheap labor from the Global Majority for data labeling and content moderation. Scale.ai employs more than 100,000 people to label AI training data, and Sama.ai moderates content for Meta through workers in East Africa.
The pay gap across the pyramid is striking. AI research scientists and ML engineers in the Global North often earn between $150,000 and $1,000,000 a year. Data workers in the Global Majority usually earn $2-5 an hour for essential AI development work. Much of the training data for OpenAI's ChatGPT was labeled by workers earning less than $2 an hour.
That's why we want to expand CDI into SINDA (Sustainable Income through Nature Data & AI). SINDA is a way to climb the pyramid. It starts with fair data collection, with transparent pricing and data sovereignty, and then builds digital skills in data analysis and AI fundamentals. It creates career paths through mentorship and specialized training in environmental AI, and eventually supports local innovation through community-led research.
SINDA's goal is a fair system where Global Majority communities take part at every level of AI development. The program aims to train our community members in advanced data analysis, set up local AI research hubs and support Indigenous-led AI research. That gives the Global Majority the skills for Level 2 funding, such as environmental monitoring contracts and scientific grants.
Taina: developing RI agents with local communities
There's a simple reason to start with local communities. Indigenous Peoples make up only about 6% of the world's population20 but protect more than 80% of its biodiversity. They're also on the front lines of the climate and biodiversity crisis. We believe regenerative intelligence in the hands of nature's stewards can create new economic and cultural opportunities for local communities and help them build resilience and protect biodiversity. Taina is our first attempt at this, an AI assistant built on Meta's open-source Llama models that helps Indigenous and local communities share knowledge.
Deploying AI raises real ethical questions. Today's AI systems are often unfair, open to attacks and hard to control. They can amplify systematic biases even when trained on balanced data. As harmful uses of AI keep appearing21, Indigenous Peoples need a voice in how AI gets built. That participation has to be sovereign and decentralized, so that it builds trust, protects data privacy and addresses concerns about data colonialism.
To develop Taina responsibly, our team helped set up an Indigenous and Local Data Council of four Indigenous and local communities around Manaus. The council helps control how local knowledge flows and where it's stored, and makes sure the benefits are shared fairly. The model worked so well that GainForest recently extended it to all our communities around the world.
In workshops with the council we found two main problems with bringing CDI and SINDA to Indigenous communities:
Current technology often doesn't fit traditional ways of sharing knowledge, such as storytelling
Communities worry about data exploitation and unequal benefit sharing (data colonialism)
In response, we co-created Taina with these features:
Support for English, Portuguese, Spanish, Bahasa and Swahili, so more people can take part
A data provenance layer that keeps shared knowledge transparent and owned by the people who shared it
Text, image and speech recognition, to fit different ways of communicating
Respectful engagement with local and Indigenous environmental knowledge
Delivery through Telegram, which is widely used in Brazil, so nobody needs to install another app
Image 2: An example conversation between a community member and Taina, GainForest's emerging RI assistant
We built Taina to put community sovereignty and data privacy first. Community members run their own instances of Taina on local machines, and each community's Telegram bot stores knowledge only for authorized users. A local image-to-text endpoint processes images, and a privately hosted Whisper API transcribes voice messages. An open-source model then handles the conversation, designed to keep it engaging and respectful with thoughtful questions and a friendly tone.
Data governance follows strict privacy rules. All data either flows through community-owned local servers or, with the community's explicit consent, through endpoints that GainForest provides. In the second case we use the data only for inference and delete it right afterwards, so communities stay in control of their knowledge.
Taina22 and its local community versions will play a big part in how we bring AI to communities, support their participation in CDI and SINDA and, eventually, use RI to scale sustainable self-governance. Taina is open and transparent. Communities can use it to build skills and as a sandbox to experiment with RI as it develops. They can learn technical ideas like constitutional AI fine-tuning, or write system prompts that match their cultural values and traditional knowledge. We believe this hands-on work lets communities shape how AI develops and keeps regenerative intelligence in line with Indigenous perspectives and needs.
Polly: developing RI agents for global climate coordination
Taina shows how regenerative intelligence can help local communities. Scaling Ostrom's principles, though, means working on coordination problems at several levels at once. So we built Polly, Taina's sister agent, who brings regenerative intelligence to international climate policy, a global coordination problem where traditional governance is falling short. GainForest and the Youth Negotiators Academy (YNA) developed Polly together. She tackles a hard problem in global environmental governance, the systemic barriers that keep youth from the Global Majority from taking full part in international climate negotiations.
The Conference of the Parties (COP) system is the world's main forum for climate action, biodiversity protection and fighting desertification23. The negotiations are technically complex, the procedures are intricate and language barriers are common. Together these often shut out voices from the Global Majority. That keeps old power imbalances in environmental decision-making in place, and it hits young negotiators especially hard, even though they bring important perspectives to climate talks.
Polly is an AI assistant that makes climate diplomacy easier to take part in. Its tools help youth negotiators to:
Understand technical climate finance and policy concepts through simpler explanations
Get past language barriers with multilingual document analysis and communication
Find relevant history and precedents quickly during time-sensitive negotiations
Draft and analyze interventions in negotiations
Polly is the SISL framework in action. Youth negotiators give us feedback at events like COP29 in Baku and UNCCD COP16 in Riyadh24, and the system improves to fit what they need. This loop keeps Polly aligned with the goals and problems of youth negotiators, and we hold it to high standards of accessibility and ethical AI use.
Image 3: YNA and GainForest present Polly during the Rio Conventions
At COP29, GainForest ran a support desk at the YNA Hotspot, where youth negotiators from 14 countries got hands-on help using Polly in their negotiations. As Emmanuel Elogima Vandi, a youth negotiator from Sierra Leone, put it: "Polly is an excellent tool. It's been incredibly helpful in navigating the first week at COP, particularly with negotiations related to the SBI and SBSTA25."
Polly is a step toward fairer and more inclusive climate negotiations. It combines intelligent systems with human expertise and judgment to close long-standing gaps in climate diplomacy, and we hope it helps the next generation of climate leaders from the Global Majority. Taina and Polly show that early regenerative intelligence can work at different scales and in different settings. To prove the approach in the real world, though, we needed a bigger test that brought technical innovation, community engagement and practical impact together. The XPRIZE Rainforest competition was that test.
Winning the XPRIZE Rainforest finals with regenerative intelligence
Five years ago, when XPRIZE announced the Rainforest competition, I saw it as our chance to prove regenerative intelligence to the world. XPRIZE has a history of driving breakthroughs with ambitious moonshot competitions. Its first prize, thirty years ago, helped start modern commercial spaceflight. The Rainforest XPRIZE was an extraordinary challenge. Teams had just 24 hours to survey 100 hectares of dense rainforest, and no person was allowed to enter the forest.
The challenge went straight at a core problem in conservation, scaling Ostrom's fourth principle, monitoring. Traditional biodiversity assessment means scientists spending months in the rainforest using invasive methods. Forests are disappearing faster than ever, and we can't document and protect species fast enough to inform policy. Automation has become essential for biodiversity monitoring, but I knew we had to go beyond the usual technology-first approach.
GainForest entered the competition early and later joined forces with ETH Zurich to form an interdisciplinary team of roboticists, naturalists, AI researchers and Indigenous scientists. From the national parks of Singapore to the Amazon rainforest, we built automated drone sampling, mesh networks, canopy raft sensors and real-time AI monitoring, all in close collaboration with local and Indigenous knowledge holders26.
Image 4: One of the team's autonomous drones deploys a canopy sensor in the Amazon rainforest during the XPRIZE Rainforest finals
The groundwork we had laid with Taina paid off during the competition. Respectful knowledge sharing had built trust, and the governance we set up with our Indigenous Data Council made real collaboration with local communities possible.
Throughout 2024 we ran eight co-learning workshops around Greater Manaus27 and asked local communities whether our technology would serve their needs. What makes me proudest is how the workshops changed over time. Indigenous team members who started as participants in our early capacity-building sessions became workshop leaders themselves. That move from participant to leader is exactly what we hoped bottom-up co-design of sociotechnical intelligent systems would look like.
Image 5: Over the years of the XPRIZE competition, we worked with Indigenous and local communities to co-design our intelligent systems
The workshops followed the SISL framework, with constant feedback between technology development and community needs. We started with basic training in AI, eDNA and drones, then asked communities to test and criticize the technology themselves. Communities that joined our frontier data collection learned to use mobile web3 wallets for transparent micropayments. During the XPRIZE, several community members became skilled Indigenous data scientists, and their communities kept ownership of their data and knowledge.
I'll never forget the moment our team, Western scientists from Europe and the US together with local and Indigenous scientists from Brazil and the Amazon, took three speedboats on the Rio Negro to the competition site. We had built more than a piece of technology. Traditional knowledge guided our autonomous drones, and our monitoring algorithms ran on data that local communities had carefully collected over months. It was a first version of a regenerative intelligent system, where technology added to human expertise instead of replacing it.
Image 6: In the last months of the XPRIZE competition, through SISL, Indigenous collaborators became Indigenous scientists who supported and led much of our technology development
In the end our team was one of the winners of the XPRIZE Rainforest, out of 298 teams worldwide. That validated our approach to regenerative intelligence. But the decision we made next meant more to me. We put our entire quarter-million-dollar prize into an endowment fund for future generations of Indigenous scientists and regenerative intelligence bridge builders28. It reflects what GainForest believes most, that lasting impact needs technology and social progress to move together.
This journey convinced me that regenerative intelligence works in practice, on hard environmental problems, and is more than a theoretical framework. By combining advanced technology with Indigenous wisdom, transparent governance and community empowerment, we showed that governing the natural commons in the future will depend on building bridges between cultures, between knowledge systems and between generations. And yes, it can even win you an XPRIZE.
Conclusion
In this essay I've looked at how regenerative intelligence could take Ostrom's principles of commons governance beyond their traditional local limits. RI is still early, but at GainForest it already works outside the lab, from communities around Manaus to UN climate negotiations.
New funding mechanisms
Our early work shows how RI can change environmental funding. Traditional mechanisms like carbon credits matter, but their high upfront costs and complex verification often shut grassroots communities out. RI-enabled mechanisms like Conservation Data Income (CDI) and SINDA are easier for communities to get into. Our research suggests they are also more robust and need less human oversight, which makes them more efficient and easier to scale than traditional approaches.
Automated verification combined with community governance lets these mechanisms run with less friction and more transparency and trust. Our Nature Guild's data council shows what's possible. We've seen communities go from having no data infrastructure to winning philanthropic funding, and one community even landed an advanced environmental monitoring contract. We believe this step-by-step approach can eventually let communities enter carbon markets while they keep their local autonomy and control.
RI agents that scale human cooperation
Taina and Polly show how machine intelligence can extend what people can do in very different settings. They support Indigenous communities in local conservation and help youth negotiators work through international climate policy, and in both places they help people cooperate in ways that weren't possible before. Our XPRIZE Rainforest win backed this up. Advanced technology combined with Indigenous wisdom solved a hard environmental problem better than either could have alone.
Positive tipping points and the race to the top
Most importantly, we're starting to see what we call "the race to the top". Communities with RI tools become more productive and earn more, and that sets an example others want to follow. It's the opposite of the race to the bottom behind so many environmental problems. Communities compete to learn, protect and regenerate their natural commons, instead of competing to extract resources the fastest.
As more communities join this loop, we expect regenerative practices to spread faster and faster. Every success opens a path for others, whether it's a community moving from basic CDI to carbon credits or an Indigenous data scientist training the next generation. Accessible funding, skills training and AI assistance remove the usual barriers, so communities can make big jumps in what they can do.
Regenerative intelligence gives us a theory of change for how humanity governs its shared resources. If we add machine intelligence to human wisdom in ways that keep people's agency and build trust, we can cooperate at scales that are out of reach today and make community bonds stronger while we do it. The answer to our biggest environmental problems may lie in combining traditional wisdom and modern technology with care, to build systems that are truly regenerative.
References
[1] Hardin, G. (1968). The Tragedy of the Commons. Science, 162(3859), 1243-1248.
[2] Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press.
[3] Simon, H. A. (1971). Designing Organizations for an Information-Rich World. In M. Greenberger (Ed.), Computers, Communications, and the Public Interest (pp. 37-72). Johns Hopkins Press.
[4] Dunbar, R. I. M. (1993). Coevolution of neocortical size, group size and language in humans. Behavioral and Brain Sciences, 16(4), 681-694.
[5] Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
[6] Jones, C. I., & Tonetti, C. (2020). Nonrivalry and the Economics of Data. American Economic Review, 110(9), 2819-2858.
[7] McKinsey & Company (2023). Carbon removals: How to scale a new gigaton industry. Sustainability Insights.
[8] ClimateWorks Foundation (2024). 2023 Annual Report: Climate Giving Trends.
[9] Statista (2024). Global Indigenous Population and Biodiversity Protection Statistics.
[10] Markets and Markets (2023). Environmental Monitoring Market Global Forecast to 2026.
[11] Amodei, D. (2024). Machines of Loving Grace. Retrieved from https://darioamodei.com/machines-of-loving-grace
[12] Bai, Y., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv preprint arXiv:2212.08073.
[13] Global Biodiversity Information Facility (2024). Global Biodiversity Data Distribution Analysis.
[14] XPRIZE Foundation (2024). Rainforest XPRIZE.
[15] Dao, D., et al. (2018). Decentralized Sustainability. Medium, GainForest. Retrieved from https://medium.com/gainforest/decentralized-sustainability-9a53223d3001
[16] World Bank. Indigenous Peoples. Retrieved from https://www.worldbank.org/en/topic/indigenouspeoples
Appendix A: Glossary
Agentic function calls: Tools and protocols that let AI systems plan and carry out a series of actions on their own while staying aligned with a community's values and goals.
Conservation Data Income (CDI): Regular payments to communities for collecting environmental data and monitoring their land. It is how most communities first join the digital commons.
Data colonialism: Taking data from communities, especially in the Global Majority, without paying them fairly or leaving them in control of it.
Data sovereignty: A community's right to control how its data, including traditional knowledge and environmental information, is collected, owned and used.
Digital commons: Tools, data and knowledge that a community maintains and governs together.
Digital trust: Confidence in digital systems and how they are run, earned through transparent operation, fair participation and clear accountability.
Dunbar's number: The cognitive limit, around 150, on how many stable relationships one person can keep. It sets the natural size of direct human cooperation.
Global Majority: People of African, Arab, Asian and Latin American descent, about 80% of the world's population. We prefer the term to "Global South" or "developing world".
Hypercerts: Blockchain-based certificates that record an impact claim in a standard format: who did the work, what impact they claim and the evidence for it.
LUCA: A decentralized marketplace that connects hypercerts with environmental data providers. It pays data collectors fairly and leaves communities in control of their data.
Nature Guild: A decentralized organization through which communities govern natural resources together and manage shared funds with transparent, programmable rules.
Ostrom's eight principles: The design principles for successful commons governance that Nobel laureate Elinor Ostrom identified:
- 1.
Clear boundaries
- 2.
Local rules
- 3.
Collective choice
- 4.
Monitoring
- 5.
Graduated sanctions
- 6.
Conflict resolution
- 7.
External recognition
- 8.
Nested systems
Polly: An AI assistant that helps youth negotiators from the Global Majority in international climate negotiations with technical analysis and policy context.
Positive tipping points: Thresholds past which a beneficial change starts reinforcing itself and speeds up.
Race to the top: A dynamic where communities compete to protect and regenerate resources, the reverse of the tragedy of the commons.
Regenerative intelligence (RI): The theory and practice of designing sociotechnical intelligent systems that preserve or enhance social capital, trust and agency while scaling governance capabilities.
Regenerative practices: Ways of working that restore and renew the resources they depend on.
Retroactive funding: Funding that rewards positive impact after it has been achieved and documented.
Self-improving sociotechnical loop (SISL): GainForest's framework for improving governance over repeated cycles of data collection, tool building and capacity building.
SINDA (Sustainable Income through Nature Data & AI): The expansion of CDI that trains communities for better-paid work higher up the data economy.
Smart contract: A contract written as code that executes itself, so agreements run automatically and in the open.
Social capital: The relationships, trust and shared norms that let a society function.
Sociotechnical systems: Systems with both social and technical parts. Solutions that work have to address both.
Splits contract: A smart contract that distributes funds automatically according to rules and weights set by community members.
Stablecoin: A cryptocurrency designed to hold a stable value, usually pegged to the US dollar.
System 1 and System 2: Daniel Kahneman's two modes of thinking. System 1 is fast, intuitive and emotional. System 2 is slower, deliberate and logical.
Taina: An AI assistant co-designed with Indigenous communities that helps them share knowledge and collect conservation data while respecting traditional knowledge systems.
Technological leapfrogging: Skipping intermediate stages of technology and adopting more advanced tools directly.
Traditional knowledge: The knowledge, innovations and practices that Indigenous peoples and local communities have built up over centuries and adapted to their culture and environment.
Tragedy of the commons: The idea that individuals acting in their own interest can together use up a shared resource, even when that hurts everyone in the long run.
Value ladder: The steps from basic data collection up to analysis and market integration, each creating more value.
Web3: A decentralized version of the internet built on blockchains, with an emphasis on user ownership, digital trust and transparent governance.